An Efficient Deep Learning Hashing Neural Network For Mobile Visual Search

The proliferation of mobile devices has resulted in a new mobile visual search application that lets users use the smartphone to sense the surrounding environment.Due to the special challenge of mobile visual search,achieving high recognition rate has become a consistent goal of related work.In the field of mobile visual retrieval,the speed of retrieval and the memory requirements of mobile devices are also crucial.In this context,this paper designs and implements a mobile vision retrieval system based on deep hash learning algorithm.In this paper,we explore the overall hash algorithm based on depth learning to construct a more powerful and real-time mobile visual search,and propose a low-parameter,low-latency and high-precision depth hashing method to construct a binary hash of mobile visual search code.First,we take advantage of the architecture of the MobileNet model to significantly reduce the delay in depth feature extraction by reducing the number of model parameters while maintaining accuracy,which is much smaller than previous models and can maintain relatively high retrieval accuracy,which facilitates the deployment of the model on the mobile for retrieval in offline mode.Next,we add a hash-like network layer to the MobileNet network model.We use supervised training model with mobile visual data.We can use the hash layer to quantify the eigenvector of the image into a binary hash code.Hamming distance is used retrieve the most similar images,which will greatly improve the retrieval speed and retrieval accuracy also have some improvement.Finally,the evaluation results show that the proposed system can outperform the most advanced precision performance in terms of retrieval accuracy(mAP).More importantly,memory consumption is much less than other deep learning models,most important,the proposed method requires only 13MB of memory to load the model and a 97.80%MAP on the visual search based location-aware data set.